URAIMSL: A Unified Retrieval-Augmented Intelligence Framework for Clinical Decision Support and Biomedical Research

Abstract

Biomedical research literature is growing far beyond the capacity of any scientist or medical practitioner to manually track it up, which happens to be exactly the scenario in which evidence-driven decision-making often breaks down. Large language models were expected to solve the problem, but unfortunately, these models still suffer from bounded context windows, from hallucinating statements and lack of persistent memory of the previous session's discussion. In this paper, we introduce URAIMSL (Unified Retrieval-Augmented Intelligence for Medical & Scientific Literature), a Retrieval-Augmented Generation (RAG) system specifically developed for decision support in medicine and biomedical research. The proposed framework combines billion-scale semantic retrieval, hierarchical multi-stage reasoning pipeline and embedding-based persistent memory module, which is purposefully detached from external document retrieval. It is this detachment that makes it possible for URAIMSL to reuse previous reasoning traces and maintain the context understanding from one session to another without compromising isolation of user data. Since the framework is not limited to any particular backbone, it can be integrated with different large language models according to specific deployment requirements. The experimental results show improvement in all key aspects of the retrieval performance (precision, contextual consistency, evidence-drivenness). These results collectively indicate that URAIMSL serves as a feasible option for advancing research and clinical intelligence in the field of AI-driven healthcare.